LungNodNet-The CNN architecture for Detection and Classification of Lung Nodules in Pulmonary CT Images

Abstract

Lung cancer detection at an early stage would be life saving. Usually it is diagnosed at a later stage which leads to increase in the mortalities. Detection of malignant lung nodules from CT images is a challenging task, given several factors that impact the detection and classification. In this work, we are proposing a convolutional neural network (CNN) based deep learning model that improves the accuracy of the nodules classification into benign and malignant types. Lung imaging database consortium-image database resource initiative (LIDC-IDRI), a publicly available lung CT scans dataset have been chosen for experiments. The proposed method come up with an approach to patchify the image to include the nodules segments of the image thus reducing the size of CT image drastically by extracting the nodule patches. Computational overhead is decreased due to the presented strategy. 6691 images containing both nodules and non-nodules are subsequently loaded into a 4-layered 2D CNN. Apparently two convolutional and two dense layers form the four layered CNN. Twenty filters having size of 5x5 is employed with relu activation function for first convolutional layer and 40 filters with size 3x3 has been specified for the second one. The model has been trained and validated on 70% and 10% respectively and tested on 20% of dataset. The verification performed on evaluation data resulted in 93.58% accuracy, 95.61% sensitivity and 90.14% specificity.